Every Model Reads a Handbook Before It Reads You
Before your message arrives, an AI has already been handed thousands of lines of instructions telling it how to respond, how to handle tasks, when to use tools, and where the limits are. That's the system prompt, and it's most of what makes one model behave differently from another.
Five Kinds of Instruction
Almost everything in a modern system prompt falls into five buckets: response style — tone, format, length; task handling for coding, writing and analysis; tool behaviour, covering when and how the model reaches for something; safety rules governing refusals and sensitive topics; and situational cases like the current date or special context. The guide notes Claude's published prompt has grown roughly ninefold in two years as more behaviour gets pinned down explicitly.
You Can Just Read Them
The useful part: you don't need a leaked copy. Anthropic publishes Claude's actual system prompts for the web and apps in its release notes, dated per model, at docs.claude.com. Read one, diff it against an older version, compare across models, and take the patterns for your own prompts.
Two honest caveats the guide includes. Those are the app prompts — the API and tool prompts aren't all published. And the unverified "leaked system prompt" repos floating around are frequently fabricated or mislabeled, so they're worth skipping in favour of the published source.
What's in the Guide
The five sections explained, why these documents keep growing, where the official prompts live, and a four-step method for studying them: read one, diff it, compare across models, lift the patterns.
The full field guide is in the PDF.
Get the Guide
Drop your email below and we'll send it straight to your inbox.